Room Impulse Response Embeddings for Speech Enhancement in Noisy and Reverberant Environments

2026-09-25

AI Summary

This study addresses the challenge of extracting room impulse response (RIR) features in noisy and reverberant environments by proposing a self-supervised learning-based RIR embedding method. Specifically, this work introduces a novel teacher-student network architecture combined with a curriculum-based multi-stage training strategy to effectively learn RIR representations from single-channel noisy reverberant speech. The learned embeddings are subsequently employed to condition speech enhancement models. Experimental results demonstrate that the proposed approach yields consistent performance gains across diverse acoustic scenarios, significantly improving speech quality metrics and reducing downstream word error rates.

Abstract

We propose a self-supervised approach for learning room impulse response (RIR) representations from single-channel noisy-reverberant speech. It consists of first training on reverberant data, then on noisy-reverberant data, and finally with a teacher-student approach, where the student learns to replicate the teacher's embeddings when given a noisy version of the reverberant input. We assess their representational capabilities by estimating acoustic room parameters from them. Conditioning a discriminative speech enhancement model on the derived embeddings yields consistent gains across all evaluated metrics, including downstream word error rate, for both reverberant and noisy-reverberant speech.

Problem

• Speech Enhancement
• Room Impulse Response
• Noisy-Reverberant Environments
• Self-supervised Learning

Innovation

• Room Impulse Response Embeddings
• Self-supervised Learning
• Speech Enhancement
• Teacher-Student Distillation
• Noisy-Reverberant Environments